 |
 |
 |
 |
 |
 |
 |
 |
 |
| NovaSearch Technology Most commercial speaker-independent speech recognition systems work by recognizing words or sequences of words in the order in which they are spoken, usually applying a grammar to define the word sequences expected. When needed, Novauris's speech recognition system can also be configured in this way. However, its unique strengths appear when it is configured differently: in the arrangement we call NovaSearch. |
 |
| NovaSearch — a powerful new way to access data NovaSearch doesn't carry out recognition at the word or sequence-of-words level, but rather identifies complete phrases from start to finish by matching them against a potentially huge inventory of possible utterances. This enables it to assemble information about what has been spoken over utterances of virtually any length and take near-optimal decisions. NovaSearch™ derives its power — power that allows direct access to 245 million items — mainly from its tight integration between speech recognition and database search. In addition, the speech recognition in NovaSearch utilizes certain kinds of phonetic information, including aspects of syllable structure, that are normally ignored in conventional speech recognition. |
|
No need for grammars... The NovaSearch method can handle large numbers of possible utterances without needing the kind of finite-state grammars that conventional recognizers typically require. In principle, every utterance can be unique in its structure (think, for example, of the first line of every song in a large music collection). Yet NovaSearch doesn't have to try to recognize each word independently: it can match the whole spoken pattern in one single process. |
|
...but with structured input the sky's the limit With unstructured utterances, NovaSearch can rapidly distinguish between up to 100,000 items. But if there is structure to be exploited, NovaSearch can handle enormously bigger sets. Novauris has demonstrated rapid, accurate identification with 245 million items! In tests with 245 million US names and addresses (artificial names but real addresses) on an ordinary laptop the correct name and address was identified in less than a second with accuracy well in excess of 99%, and less than 100 MB of memory was needed. |
|
NovaSearch can cope with variation in inputs... NovaSearch technology can allow variations in the way that users choose to specify items. Using patent-pending techniques, selected words can be inserted, omitted or reordered. For example, the name-and-address application can be arranged to accept spoken input in which components such as the state are left out or the order of the given and family name are reversed, or a title such as "Mr" or "Ms" is added. |
|
...and even allow spontaneous, informal spoken control On smaller tasks, for instance a few thousand basic commands for a robot, the amount of variation allowed can cover virtually any way of formulating the command: Get me a coffee, Please get me a coffee, I need a cup of coffee, A coffee please, Fetch me a cup of coffee, Can I have a coffee... In summary, NovaSearch offers a unique way of speech enabling a wide range of tasks ranging from access to hundreds of millions of items such as names and addresses to command and control applications needing an informal, highly flexible interface. |